AI Has a Subsidization Problem  We want to be transparent that we'remaking changes to Gemini CLI that mayimpact your workflow, including morerobust detection of policy-violating usecases, prioritizing traffic for certaintypes of accounts and restricting GeminiPro models for free tier users. Important update to GitHub Copilot forstudents. Your complimentary copilotaccess will be managed under a newGitHub Copilot student plan. As part ofthis transition, however, some premiummodels including GPT 5.4 and ClaudeOpus and Sonnet will no longer beavailable for self-selection under theGitHub Copilot student plan. Open Code1.3.0 will no longer load the Claude Maxplugin. We did our best to convinceAnthropic to support developer choice,but they sent lawyers. Hi, Google team. I'm a Google AI pro subscriber, and Ibelieve there's a bug or undocumentedchange in how quotas are being appliedto my account. We should receive highgenerous quotas refreshed every 5 hours. With higher weekly rate limits than free users,however, my current anti-gravity UI showing much longer cooldowns. Itlooks like the subsidization era mightbe coming to an end faster than weexpected. I'm pretty surprised, to be honest. I did not think Googlewould be the first company to really tapout here and start cutting off the freeand subsidized access to users who arestill hyped on Gemini stuff for somereason. It's kind of crazy to see howquickly the tides have turned here. Fromgiving out thousands of dollars ofinference for hundreds of dollars tolocking out free users entirely andmassively restricting what student planusers can have on things like GitHubCopilot. This is the start of the end. But in order to understand andappreciate the end, we probably need tobetter understand the start, too. How wegot here, what this all means, and whatthe future of the economics of AIdevelopment stuff is. Are theseCompanies even going to exist in a fewyears? I have no idea. Thankfully, noneof them are paying me for any of mycoverage here. So, I'm going to do mybest to cover this all in an unbiasedway and help you better understand wherethe tools and the subscriptions we'repaying for now are going. But in orderto pay for them, I need some money. So,we're going to do a quick break fortoday's sponsor. Sometimes the popularsolution isn't very good. And I wouldconfidently say that's the case withGitHub Actions. The fact that we runmost of our CI through them is pathetic. Even if you're using something awesomelike today's sponsor, Dep Depot, theymake your actions way faster and alsoyour Docker builds. But they are kind ofhampered by what GitHub allows becauseGitHub Actions, to be frank, suck. If onlysomeone who knew how to make actionsbetter was to take the time to makebetter action. Oh, what's that on myscreen? Now available. DepotCI. Oh,Depot fixed GitHub actions by killingGitHub actions. I really like how theyframed this year. They did everythingthey could to make GitHub actions asgood as possible, but there was always aceiling they had to hit. And now they'vehit it. Turns out they can only improveabout 30% of the system. But the other70% was outside of their control. Thatmakes it slower. That makes it harder tointerface with. That makes it impossiblefor your agents to do anything. And I'msure we've all had to copy paste errorsfrom our CI back over to our agents. What if there was a CLI they could useto actually get that information? Whatif there was more reliable ways to SSHinto the boxes that are running yourstuff so you can actually figure outwhat's going wrong? What if there weredashboards that gave you actually usefulinformation and skills you could handyour agents to take advantage of all ofthis? What if GitHub actions werereinvented today by people who knew whatThey were doing and wanted to enableagents to be more successful at buildingin your real code bases? And what if youcan migrate everything over with twocommands? Yes, it's actually that easyto get started. It's already way fasterand better to use, not just for yourselfas a developer, but for your agents,too. Stop fighting GitHub and get backto work at soyv.link/depo. There's no such thing as free compute.And I want to be very, very strict aboutthis fact. People seem to think that AIis just free once the models arecreated. And they're not. Just runninglocal models on my own GPUs in myapartment costs me hundreds of dollars amonth in electricity alone. These thingsdo cost money. And if you're giving themout for free, there has to be a reasonfor that. It could be that you're makingmoney on ads. It could be that you'retrying to sway a customer with a trial. It could be that you're collecting a tonof data. And chances are it's someCombination of all of those things, inorder to make this economically viable,but we're not getting to a place wheregiving out free inference in generalmakes a lot of sense. And this is atheory we've had for a while. The priceof frontier intelligence has been goingdown somewhat consistently. This is achart that shows how the cost ofdifferent models dropped from 2023onwards, and things like 40 and 40 miniwere upwards of 10 times cheaper thanthe equivalent of GPT-4 32K back in early 2023. And it seemed like this was goingto stay the trend, especially withopen weight models coming out and beingeven cheaper. But there was something wedidn't expect. We didn't expect theamount of inference we are doing to goup exponentially. If you were to do asimilar chart here for the number oftokens a model generated in order torespond to a question in 2023 versusnow, the amount of tokens being used hasgone up at least 10x because now we'rePrompting in things like our code bases. And the models aren't just oneshotting a response, they're doing tool calls,which means more additional responses,more additional steps being done to gocollect data and then another chateffectively happening for each step. Reasoning tokensmassively increasing costs here as well.Massively increasing the amount oftokens being generated, the amount ofmoney that we are spending, not just onAI in general because we're using AImore, but for a given prompt, has morethan counteracted the amount. The costshave gone down, but more importantly, theamount we need the GPUs to be doingthings for has gone up. If one chatmessage used to generate 200 tokens, andnow it generates up to 200,000 tokens,that's also more time being spent on theGPU, and if you're using the GPU, thensomebody else isn't, and that costs moneyinherently. So while intelligence isgetting cheaper in the sense of per. Token sometimes, in some cases, it's notenough, and the demand for doing theselong generations that take more GPUs andmore time on the GPUs is consistentlygoing up. So where does Gemini come inhere? Well, it comes in because of thesubsidization wars. When AI models firstkind of became a thing, we were tryingto figure out how to charge for thembecause users at the time, and even now,to an extent, don't really understandtokens, and the idea of building per setsof characters didn't make sense to them. So many products, including ourselves,with T3 chat, decided to bill based onthe number of messages. That's aproblem. If I go to a service like T3Chat and ask it, what's 2 plus 2? Youget an answer generated very quicklywith not too many tokens, only 11 tokensto answer this. But if I ask it to write15 poems about React, I won'tdo 15. I'll do five. Be conservativehere. This message was 434 tokensgenerated. That's 40x more tokens for aRoughly similar question length becausethe thing I'm asking for is just morecomplex here. And the answer is going tobe more. And if you think I kind ofexaggerated this gap here with the 40xexample, I didn't. In real-worldexperience, I could say that the gapbetween the bottom 10% most token usedmessages and the top 10% most tokens usedin a message is closer to a 400x gap. Sosome messages cost literally 0.001. Other messages can cost multipledollars. And yes, you can send a messageon T3 Chat for eight bucks a month wherethe individual message cost us on theAPI a dollar, an eighth of your sub withone prompt. It's kind of insane, butthat's just how the models work and howthese things naturally can cost morewhen more is happening. At the sametime, these business customers aregrowing their needs. The desire forthese companies to compete with eachother is going up too as they fight forend users. Open has pretty much won theEnd user right now with the $20 a monthChashibbt subscription that makes up 70% oftheir revenue. They're killing it there. I don't think I've ever met a person whoshut down their $20 a month Chashibbtsubscription to go subscribe to $20 amonth of Gemini. No one does that. Theclosest I've seen is people who use thefree tier of Gemini and realize theydon't really need Chashibbt anymore. So,they go over and cancel it. And I getthat. I understand that. And this iskind of the bet Google wants to make iscan they make intelligence good enoughand cheap enough that every user willjust start using Google's intelligenceinstead, eventually killing all thecompetition and then Google owns thewhole AI world, and they can chargewhatever they want for the companiesdoing the categorization. The problem isthat we're all fighting over the sameGPUs. Every one of these companies hasan amount of infrastructure they can useto serve requests for their business. Customers, for their research labs,generating things with AI and makingtheir own models and making their ownwhatever they're doing, and the end userswho are paying the whatever a month togo use the models in products likeanti-gravity or T3 code or just using itin their chat interface directly. Andsince Google in particular desperatelywants to steal customers off of OpenAIand off of Anthropic, they've been waytoo generous with subsidizing usage intheir products. I cannot tell you howmany times I have seen comments fromlower-income people in my content, be it Twitter, Twitch, YouTube, or wherever,saying that I need to stop being so meanto anti-gravity because it's the bestoption for poor people who can't afforda subscription for something better. Towhich I would usually bite my tonguebecause I know the reality here. Thosecustomers aren't valuable enough to getfree inference. There is no world inwhich that continues. And we are now atThe end of that. I had a feeling itwould come, but I honestly thoughtGoogle's billions upon billions ofdollars would sustain this a bit longer. Google got users through giving out freeinference, but that meant the majorityof their users were using freeinference, which meant that they wereeating a ton of costs and had a lot ofGPUs that were being used by users whowould never spend actual money when theycould be used by people who arecurrently spending actual money. Andthis has affected us. We've had timeswhere people are using paid things on T3Chat, where we're paying API prices, andthey send a message, and 3.1 just failsto respond because it's overloaded. Oreven I, on the $250 a month Google plan,that I subscribed to forever ago forcontent and probably should unsubscribe,but it's fun to have it there to [ \_\_ ]about things like this. I couldn't usethe new Gemini 3.1 Pro model when itdropped in the $250 a month subscription. And the Gemini CLI. And when asked, theycited this as a capacity issue becausethey're giving it away for free too much. There is no such thing as freeinference. Either you're the product andyou won't be for long or somethingillegal is happening. So, if you werepicking these tools because they're freeand you can't afford it and you didn'tuse the two months of thousands ofdollars of free inference, mind you,like you're getting thousands of dollarsof compute for free during that window. If that wasn't enough for you to find away to make 20 bucks a month, that'syour problem, not Google's. Seriously,like this, this was just going tohappen. I am surprised it did so quick. But now there is zero reason to be usinganti-gravity or the Gemini CLI becausethey are not subsidizing anywhere nearas hard. They are kicking free users offof the pro models entirely. And as muchas I am a Gemini Flash simp to anextent, they are not good enough forDay-to-day code, the idea of usingsomething like Gemini 3 Flash on thefree tier or Gemini CLI for doing realcode makes me feel somewhat sick. It'slike using a toddler's like toy keyboardto try and build a real application orlike using a FisherPrice karaoke machineto record an album. Like it's just it'snot there. It feels silly. The craziestline here though is around directcontrol. For direct control over yourown quotas and billing, use your ownpaid API key from AI Studio orVert.Ex.ai. This is them saying directlyon the official Gemini CLI account,we're going to cut down your usage onthe paid subscriptions. So if you wantmore usage, just go use an API key. That's crazy. That's insane. They'recutting you guys off hard. So how do weget here? Why would these companies alldo free and subsidized inference? Itseems to just bring drama. Like rememberthe drama in July of last year whencursor switched away from a per messageTo a billing on usage model and everyonefreaked the [ \_\_ ] out. It seems likethese things just cause problems. Well,the first reason you would do this isyou can make money elsewhere. Probablyads. If you have the ability to makeenough money on ads to cover theinference being done, awesome. But Ithink people massively overestimate howmuch money you can make on ads. One ofmy favorite ways to demonstrate howlittle money there is in ads is myself. My YouTube channel is pretty popular. Weget around 3 million views per monthfrom experienced developers. So, our CPMis relatively high. In the last 28 days,my channel got 465,500hours watched. So with 672 hours in the28 days, that puts us at 693hours per hour. So every time an hourpasses, 693 hours of my content arewatched. For those who have beensubscribed and watching for a long time,we appreciate you immensely. Thesenumbers are insane. They're still hard. For me to just process, that we'renearly at a thousand hours watched perhour of time, that passes is justmind-boggling. That at any given moment,on average, 600 to 700 people arewatching Theo talk is crazy. So, howmuch money do you think I make off ofthat per month? Remember, experienceddevelopers, high CPMs, because thesedevelopers are experienced. How muchmoney do you think I get off of YouTubeads per month? And again, do not answerchat if you already know the answer. Geta number in your head. I'm going to hopover to Twitch chat so we see whatpeople here think. Okay, we're seeing40,000, 20,000, 10K to 20K, 500 bucksfrom short used to work at Twitch,you're cheating. The answer, it's about$9,000 a month. It varies between 5 to 9depending on a lot of different things.I want to emphasize the absurdity here.This is 20,000 days of my contentwatched in 28 days. So, 20,000 days of mycontent being watched is worth about. $9,000 with ads. That's not even enoughto pay my editor. Ads on YouTube are nota real way to make money, even at thesetypes of large-scale operations. And thisis revenue that is being split with me.So, Google's complete ad revenue here isprobably double this. It's believed thatit's a 50-50 split. There's no way forus to truly know for a fact that's thecase, but I take them at their word. It probably is. So, Google, off of mycontent, theoretically made upwards of$18,000total in that time. So, if you think adscan cover the costs of inference, you'revery, very dumb. Like, that's all I cansay there. If we reframe this to bedollars per view, we are at .28per view. Do you understand how smallthat is? And remember, an individualmessage can cost a dollar or more. So,if you think a little ad that makes afraction of a cent underneath yourprompt is enough to justify the promptbeing free, I have a whole bunch ofBridges down the street I would love foryou to take a look at. Seriously though,like there is no money in advertising, at least on an individual level. Peoplesee companies like Google pulling inbillions for ads because they're doingit at the scale of billions. To make abillion dollars on ads, first you mustsend a trillion ads out. And if each ofthose instances could cost a dollar,you're just going to lose money. Sothere's no world in which.3 cent ads aregoing to cover multiple dollars ofprompts. The world just does not workthat way. So if you think ads are enoughof a way to fully subsidize the cost ofthese things, that you should be able to get free inference because advertising,you do not know how any of this works. Ijust want to make sure that the scale ofthis is clear because people seem tothink ads make a lot more money thanthey do. This is also why we’re sothankful to have awesome sponsors onthis channel because they make thisWork. If it wasn't for sponsors, I wouldnot be able to do this because revenuewe get from people paying me on Twitchand from ads on YouTube is just barelyenough to cover one of my team members. Far from the entire team that we have torun the channel with. It's certainly farfrom me making a scent on it. In fact,I'm so thankful for my sponsors, I'mgoing to tell you about one right now. AI's gotten really good at buildingapps. Well, web apps at least. Once itcomes to mobile, things get a lotmessier. Even just trying to release theapps can be obnoxious. You want to takeadvantage of the cool things your phonecan do, be it the AR side or the cameraor all the other magical stuff embeddedin the Swift bindings for your iPhone,you're kind of out of luck. Well, youwere before Ror Max dropped because thisis crazy. I'm actually like genuinelyblown away. I've been setting it up toplay with for this ad and I think I'maddicted. It's hilariously easy to makeReal applications with this. The one Imade is a bit goofy, and you might noticethat you can't really see it here. That's because I have it plugged into myphone. Ror now, it ships a Mac app so thatyou can link your phone and yourcomputer together as you use cloud codein the background, by the way, in order tobuild real native iPhone applications. This is a real Swift app that was builtand it wasn't even built on my machine.It was built in the cloud, and then theyuse their pipeline to put it on my phone,when I'm working on it with my computer. The task I gave it here was not trivial.I asked it to build an app that uses theAR features and the camera on my phoneto put a hat on my head. And it actuallyworks. Like what? It's kind of crazythat something like this is able to bedone just playing around with a web applike this. You might notice the hat'slike in front of my head. That's notquite right. I asked it to fix that.Let's see if it succeeded. You seeThere's a little building button here inthe corner, and I can tell it to rebuildand update the app on my phone directly. I also just want to say the onboardingis great. All of these types of littlethings that are really annoying to getright, it just tells you how to dealwith them. And I've never had such asmooth setup doing an iOS thing in mylife. Even official Apple stuff issignificantly harder than this was. Holy[ \_\_ ] it actually worked. What? Likethese are the types of things that justwere not realistic for people to do ifyou didn't know iOS. Well, you can buildanything with this. As far as I cantell, they've shown off everything fromsubway surfers clones to flightmanagement apps to full-on Pokémon Goclones. You can use all of the thingsyour iPhone is capable of doing, all ofthe APIs, all of the stuff that wouldnormally be a blocker. But the biggestdeal by far, two clicks to App Store. Iam blown away with Rorm Max, and I betYou will be, too. Check it out now atsoyv.link/roor. While sponsors can make me money, ads inthis generic fashion cannot make enoughmoney to cover inference. Hell, even mysponsors would not be willing to cover adollar plus per prompt for all of youguys to prompt for free. Just beingrealistic here. So what other reasonswould somebody want to do free orsubsidized inference? Another one that Isee people talk about a lot and is realsome amount is the data. We have nowseen in various different scenarios thathaving chat histories with LMS that haveverified good results at the end isreally powerful for training new models.You will never be able to make a modelsmarter than the one that generated theoriginal data you're training on, butyou can get goddamn close and you can doit a hell of a lot cheaper. This is whyEnthropic is so scared of all theseChinese labs getting access to promptsand other data from enthropic models That they can use to do their owntraining. This is also why there havebeen some sketchy things that I’ve heardabout and wish I had better examples toreference and show you guys, but I, Iknow from a good friend that there weresome Chinese companies offering cheapersubscriptions for Claude code that wouldrequire you to use their weird clawrapper thing and it would still useanthropic models for a much cheaperprice, like 10 bucks a month, I think. Andthe reason they did that is they wantedto middleman every request and collectall of the inputs and outputs so theycould use that to go train things. Atleast that’s my assumption because Idon’t have any of this on the record. Itwas very illegal. They’re also probablyusing all sorts of questionable methodsto get that inference and all of that. But like that type of bootlegging isvaluable because the data is sovaluable. I would even argue that in thecase of something like cursor, theyBenefit greatly from this. Obviously,don't benefit enough to give you freeinference because, to be frank, the peoplewho are unwilling to pay are alsoworking on jack [ \_\_ ] [ \_\_ ] in terms ofuseful stuff. So, if you can't afford a$20 a month sub, your inference is worthless than the average person. If youdon't have a job at a real company,working in a real codebase, and you can'tafford a cursor sub because of that, wedon't want your data. It's not useful. But, if you are working at a company andyou're just trying to min-max what youget per dollar, cursor can justifysubsidizing some amount because, if theycan get the data from your work and fromyour prompting, they can use that totrain their own better, cheaper modelsthat they can then use to lower theirown costs. That, all said, the data isnot valuable enough to make this worthgiving out for free. It just helpscursor eat a little more cost than theywould have otherwise been willing orAble to do. And here's where we get tothe real reason the companies are doingthis. Stealing customers. This is adesperate attempt from the companiesthat have the money to afford it toconvince people to use their thing. Andgenerally speaking, there's two reasonssomebody will change the thing they areusing. It's because there's somethingelse that's better or there's somethingelse that's good enough and cheaper. Those are the two reasons people willtry something new. In a world whereeverything around us is changing, evenidentifying what is better has gottenreally hard. And at the same time, a lotof these things have gotten moreexpensive. Like I went from having liketwo $20 a month subs at the start oflast year to three $200 plus dollar amonth subs. That happened real quick. That means getting customers is harderbecause I'm not going to just go blindlysign up for your $200 a month plan. Ineed to know it's better enough for myExisting $200 a month plan to even consider it or look at it or care. Butif you offer it to me for free, then Imight take a look. And if your thing isgood enough and free, I might go cancelthe things I'm paying for and stick withyours. or if it's cheap enough, justmove to it and be happy. If you can't bebetter than your competition or you'restruggling to communicate the ways inwhich you're better than yourcompetition, price becomes the easiestway to communicate. And I'm definitelynot speaking from experience here. Putting out the $8 a month chat app, T3Chat, definitely not something I wouldknow way too much about. If you can'tconvince people you're better, convincethem you're cheaper cuz that's easy. Andthen when they move over, they willdiscover how much better what you builtis if it's better. Or they won't. Sowhile going cheap or free can give you ahuge boost in potential new users, ifyou're not good enough, you will notKeep them, and you will just be wastingmoney on the people who are leftover. Soif we have this range of people and, onthe right side, we have billionaires, andon the left side, we have 12-year-olds,learning to code, the spend potentialfor these groups is massively different. And if you can make a product that isbeloved by billionaires, you shouldprobably do that because you can make alot of money. But there's a linesomewhere in this spectrum that is, I'lltry it if it is free. And this line isthe point at which people will be moreinterested in a thing if it's given outfor free. This line's probably furtherto the right here. Even like some of therich people I know are much moreinterested in trying a free thing than apaid thing. So, there's a line somewherehere. But there's also an importantrange to identify here, which is therange of, I will pay for the betterthing. So, if you make your thing freeand you want to get paying users longTerm, this range here is a gold mine. The range between I'll try it if it'sfree and I'll pay for it if it's better.If you can get them in because it's free andthen they stick with it because it's better,and they start spending money, awesome.You can get a lot of good users fromthis bucket. I would argue somethinglike GitHub has done this really well,where GitHub is a thing we all usebecause it's free and we see everyoneusing it and then when we go to a job, wewant to use it there. We start paying,and they make a lot of money. But here'sthe problem. We are expecting that usersin this area will be happy enough withthe product that they become payingusers. But what if the product isn'tvery good? Well, then we hit a differentline that exists in this section. I canonly use it if it's free. And this lineis a death sentence. Everything left ofhere is a suicide mission. If your onlyusers are people who are using yourthing because they can't afford to use. The better thing, you're [ \_\_ ] You'reabsolutely [ \_\_ ] You will neverrecover from this position. The onlyreason so many of these companiesoffered free tiers in the past, likepre-AI, is because everyone would try itif it's free and enough people wouldstart paying, then it was worth it. Butif all the people over this line arehappy and aren't going to try yourthing, all you're getting by making itfree is a bunch of [ \_\_ ] users that aregoing to cost you a ton of money andnever convert into real paying users. Bybeing [ \_\_ ] and being free, Google haseffectively filtered down theanti-gravity user base to be earlyprogrammers who don't really know whatthey're doing and won't be valuable andwon't spend money or really experiencedprogrammers who won't spend moneyeither. Looking at you Linus Tovals. Isaw you vibe coding in anti-gravity for some reason because it was free. Butthat's the reality of this. Most ofThese companies are doing the free thingbecause they're hoping someone likeLionus Torvos will try it because it'sfree. But in reality, what they end upwith is all of these people who are onlyusing it because they can't affordanything else. And those users are goingto waste a bunch of your GPUs and energyand time and support. Especially thesecheap and free users are so expensivesupport-wise because they have so manyneeds. There's the classic meme of the$100 client will ask for all of yourtime and attention. The $1,000 clientsays looks good. That's a very realthing. The users in this range aren't asexpensive as the paying users or lessexpensive. They are often moreexpensive. They have higherexpectations. They are more annoying. And the likelihood they ever make youmoney isn't near zero. It is zero. AndGoogle's inadvertently put themselves ina position where the majority of theusers of their code live in thisCursed corner. And that's why Google hasto cut things off. But I want to comparethis to other businesses. I want tocompare this to something like ClaudeCode. There's another line somewhere inthis that is I can afford $200 a month. That doesn't mean you're willing to justgo sign up and use it for that, but youare capable of spending that much money. So maybe you tried Claude on the freetier on the website. Maybe you got oneof those guest pass things and you'retrying out Cloud Code. You’re like,"Yeah, this is pretty impressive. I cansee myself paying for it." So, you bumpto the $20 tier. You use all your usage.You're like, "Yeah, this is useful. I'llpay a bit more." You bump up again andthen eventually you're suddenly at the$200 a month tier. You then see thatOpenAI put out a new model and youreally want to use it. You don't want tohave a second $200 a month sub though.So, instead, you go try out Cursorbecause Cursor has a $200 a month plan. But it has Claude and it has OpenAImodels in it. You were barely even usinghalf your quad sub anyway, so you'll befine. So you move over to Cursor and youhit your limits in the first 3 days andyou're like, "What the [ \_\_ ] I didn'teven get close to hitting my limits whenI was using the Claude plan. Why is thesame money I'm spending on Cursorhitting limits so fast? Did I use thewrong model? Did I do something bad? IsCursor just overcharging cuz they'regreedy?" No, it's actually quite theopposite. Turns out that that $200 amonth plan can be used for up to $5,000of compute per month. So why would theyever do this? Let's think about someuser archetypes. Let's focus on that12-year-old again. 12-year-old could useup to 15 in inference. They pay $0. Theywill pay zero over the lifetime. This useris not going to make the company money. And there are millions of them. This isnot a very appealing customer breakdown.This is not the person I would want toHave as my user. So, what about some42-year-old dev? They could use up to$5,000 in inference. They pay 200 amonth. They might pay 1 million plus overlifetime or get their company to. Andthere's a few hundred thousand at most ofthese people. So, on one hand, this42-year-old dev is costing Anthropic waymore than this 12-year-old is. The12-year-old's costing 15 over zero. The42-year-old dev is costing five grand over200. So, they're costing Anthropic theoretically up to $4,800. And the12-year-old's only costing them 15. Butlet's think about this longer term. Whatif the cost of inference goes down by10x? This free user goes down to $1.50,but that's not enough to be covered onads. That's still not profitable. Andthey'll never spend money. So, this userstill sucks. But that $5,000 ofinference is suddenly $500. Obviouslystill not profitable. They're stilllosing 300 bucks on that. But if theprice gets 10x cheaper, it gets a hellOf a lot closer. It also is worth notingthat most users don't come close tousing all of that inference. So if yourtop 1% user is using the full amountthey're allowed, but the mean usersare using less than half, you're now profitable with this change easily. Butif it gets 100 times cheaper, the12-year-old still cost 15, which mightbe almost enough to cover them with ads,or at least close. But it also meansthat Anthropics cost is 50 bucks and nowthey're getting a 4x profit on a toptier user. This is just one of thedifferent dimensions that has to beconsidered here. The more important oneis that if you give me a $200 a monthplan, I go use it personally. I reallylike it. And then I go to my company andsay, "Hey, we should use this here."They subscribe the entire engineeringorg to it. 20% of that org actually usesit. And 10% of them use it heavily. That's 80% of my spend going to peoplebarely using the thing. Just paying freeMoney to Anthropic. I'm kind of doingthis myself because I’m maintaining the$200-a-month sub, and I barely use thecloud code one because I’m too busyusing things like Cursor and Codec CLI,mostly Codec CLI, so I’m barelyusing my inference, and that helpssubsidize the cost for others. But ifall of your users spend $0, they can’tsubsidize each other. And to those whoare saying, "But what about the kidswhen they grow up? They might use thething when they’re adults. It’s not thatthey’re going to be 12-year-old and poorforever. Yeah, they’re going to becomeadults and they’re going to move to agood thing. If you were right about thisassumption, then we’d have a bunch of22-year-old Genzers who are usingChromebooks as their daily drivercomputer. You’re just wrong. When thekids grow up, they get better things. And the only reason it makes economicsense to make things for the poor kidsis if you can make money off of them. And you can't. And you're not going tomake that money later. You're justwrong. And thankfully, only one companywas stupid enough to think otherwise.And it's everyone's favorite Google. AndGoogle has now realized how badly they[ \_\_ ] themselves by doing this. Andthey're trying to lock it down. I willsay with near 100% certainty that thistype of subsidization has to die. Thereis no reason this ever should haveexisted. It was bad math on some PM's[ \_\_ ] plate that resulted in themgiving out way too much inference forfree or really cheap in a desperateattempt to get somebody to care aboutGoogle's developer products. And theyfailed outright. This type ofsubsidization is a little bit moreinteresting. On one hand, it is absurd. The amount of infants we're getting hereis crazy. This can't last. They can'tpossibly afford to do this long term.But on the other hand, the amountthey're subsidizing is going up. According to the same reason I citedearlier, which was Curser doing theirown internal testing, last year, youcould only use up to $2,000 a month forthe $200 a month plan. So, it was a 10xsubsidization. It is now a 25x sincethen. Anthropographic is increasing how muchyou're getting in terms of a discountwhen you use their models. And there's alot of reasons and a lot of incentivesthey have to do that. But the biggest isthat they want to get people off ofcursor where every model can be used andonto Claude code where only Anthropographicmodels can be used. They have a verystrong economic incentive to do this, and they are going to keep being [ \_\_ ]because of it. It's the same reasonthey're being [ \_\_ ] to DAX and opencode. There's a lot of layers to thisone, but one of them is that if you useother harnesses, it's easier to switchto other models. So you're less stuck inthe Anthropographic ecosystem. And it does notmake sense for them to give you $5,000Of inference for 200 bucks a month ifyou've moved off before inference getscheaper. These subsidies only make senseif they make you a lifetime customer. Soanything that risks losing you as alifetime customer is a massive risk. They're not interested in taking. Sothat's why Anthropic is working so hardto kick you out of tools that give youchoice. And this is the only reason I'mscared about the cloud implementation inT3 code because according to their termsof service, their SDKs, their ownguidelines, the advice I've gotten fromexisting employees and people online,and most importantly, according toClaude, our implementation is totallyfine and safe. But the economic realityof Anthropic not wanting an option thatgives you choices because then theirsubsidies don't guarantee profit long term. That's a real risk. So the onlyreason T3 Code would end up not havinggood cloud code support over time isbecause Anthropic specifically doesn'tWant you to have options. So why doesOpenAI do subsidies then? I thinkthey're getting like the opposite effectbecause some people, including thisparticular dev, have switched off ofClaude because they're tired of themodel just being shitty in favor ofsomething like OpenAI's models. This islargely because of their growth phase. In fact, I can show you. Try the Codexapp with 2x rate limits until April 2nd.By the way, it is my understanding thatthese limits apply in all Codexservices, not just the Codex app. Butthe 2x rate limits they’re giving youhere are because they are trying reallyhard to pull people over from otheroptions. Because Codex is significantlyless popular than Cloud Code still tothis day, they had to do whatever theycould to try and pull users over. Andthese rate limits are too generous. Ihaven't even hit more than 30% usagewhen I was trying. One of my employeesgot close to 90% once, and it was Julius. And he was trying [ \_\_ ] hard to do it.And then there's the reset button thatTibo keeps hitting. That's a wholeseparate set of jokes. The reasonopening I can do this and justify it isbecause they have so many subscribers onthe $20 a month tier doing nothing thatit more than justifies this. It as aland grab, as an opportunity to try andtake some ground from Claude code isworth the money here. But again, to goback to my list, this is all part ofthree. The only strategy that makes realsense. The reason to give away free andsubsidized inference is to steal customers. It will not create newcustomers. It will not grow the pie. Thereason you do this is to take more ofthe pie away from someone else. Andthat's why OpenAI is doing a lot ofdifferent things now. Not only is thiswhy they're giving us the temporary 2xrate limit, this is also why they'reworking with people like open code andlike us with T3 code to make sure weHave everything we need to integratebecause they don't want to be the onlyoption. They want to be the best option,because right now they are the option with 30% of the market at most. But ifthey give everybody a list of options,and theirs is the best value and thebest option, they will start to win moreand more. And this is going very wellfor them so far. Whereas Entropic, whilehaving incentive to subsidize to keeptheir holds, they do not have incentivesto keep offering their product nearothers. This is why you'll hearengineers working on Codeex talk aboutClaude meaningfully, but you'll reallystruggle to get a Claude engineer toeven mention Codeex. It is just theunique shape of reality distortion bubbles they both live in. But it allcomes down to where they are in themarket. And if I'm being honest, as muchas I've loved working with OpenAI and asgreat as the people I'm friends withthere have been and just are pleasant toWork with, I would suspect that if theyget over 50% market share, that theirtune will change quickly. Much like ithas for Anthropic, who went fromreasonable, somewhat generous to ifyou're not using cloud code, go [ \_\_ ]yourself. And that happened really fast. So where does Google fit in all here?They don't because they're clueless.Google as a business is desperatelytrying to get real AI customers. Andthey're doing that by forming dozens ofdifferent teams across the company withno communication and understanding ofeach other that are fighting over GPUsand allocation. And I've known peoplewho work on dev tools at Google thatcould not convince Google to let themadd Gemini Pro 3.1 because they wantedto reserve it to give it to the freeusers in anti-gravity. Yeah, Google subsidized so hard they[ \_\_ ] over their own developers. Andit's truly insane how much things are onfire there as a result. Where EnthropicDoes these types of things with a veryfocused strategy, Google's doing thembecause they're incompetent. Theyaccidentally put themselves in a badposition. So, while I'm doing my best togive you guys this framing that can helpyou understand all of these companiesand why they work the way they work,none of these framings will ever workfor Google because Google is not workingbased on what makes sense. Google isworking based on how Google works, whichis not particularly well. I've beenplanning on doing the Gemini is badvibes video for a while now. I don'teven know if I'm going to bother becauseI don't think anyone cares. The only people I knew who were using the Geminimodels and were using things likeanti-gravity heavily were the people whocouldn't afford other things. And nowthat they have officially been bootedfrom doing such, I suspect no one willbe asking me going forward. Hopefully,this helps you in your understanding ofHow these business incentives work, andmaybe you'll look at your $200 a monthsub with a little bit less scrutiny. Orat the very least, you'll be a little bitmore generous to services like us, withT3 chat that have to pay the full pricewhen nobody else does. It's a great timeto be a user of these subscriptions, butit is a terrible time to be a competitor,trying to fight them. That's why all these companies are doing it. That's whysmall startups like ourselves get kindof screwed. And that's why you, adeveloper building with these tools,have an incredible opportunity to doawesome things. Seriously, if you canafford one of these subs and you don'thave it yet, you really should go findone. Even if that means not usingsomething like what we built with T3,because they are such a high-valuereturn if you can actually use them. Thatit's one of your last chances to getthis type of free value in life. Youshould take advantage of it while youCan, but expect it to disappear any day as we're starting to see now. And hopefully, prices will get cheap enough that we don't need this long-term.Until then, though, I'm going to do my best to max out my $200 a month subs.You should too.And if you're looking for a place to do it, you should try out T3 Code.We offer all of these things in an open source rapper.Anyways, I have nothing else on this one.I'm curious how y'all feel., AI Has a Subsidization Problem, YouTube, 01kmjntmcx0ndw05w39m0m3yz3.